An efficient shrinkage bootstrap bias estimator for smooth functions of sample means (Q2488383)

For a multivariate i.i.d. sample with sample mean \(\bar X\) and (unknown) theoretical mean \(\mu\), a parameter \(\vartheta=g(\mu)\) is considered where \(g\) is a smooth function. The estimate for \(\vartheta\) is \(T=g(\bar X)\) . The usual bootstrap bias estimator for \(T\) is \(\widehat{bias}=B^{-1}\sum_{b=1}^B U^*_b\), where \(U^*_b=g(\bar X^*_b)-g(\bar X)\), \(X^*_b\) are bootstrap resamples of the data. The authors propose to use shrinked \(\tilde X_b^*=\bar X+\delta(\bar X_b^*-\bar X)\) istead of \(\bar X_b^*\) (\(\delta\) being a fixed number \(0<\delta<1\)). The corresponding bias estimator is \(\widetilde{bias}\). Conditions are given under which \(\widetilde{bias}\) is (asymptotically) more efficient then \(\widehat{bias}\). Results of simulations are presented.